Abstract: In order to quickly obtain knowledge information from tobacco-related scientific and technological literatures, an interactive iterative learning method for tobacco knowledge entity annotation and recognition was used to construct a corpus for annotating tobacco-related texts. A text annotation specification suitable for the field of tobacco was designed, and the BERT+CRF (Bidirectional Encoder Representations from Transformers + Conditional Random Field) deep learning network model was used to recognize and pre-annotate tobacco named entities. Combined with manual proofreading, the size of the original corpus was increased and the performance of the model was optimized. The results showed that the consistency F_1an of the corpus annotation reached 92.4%. The BERT+CRF model has better recognition ability than those of commonly used CRF, BiLSTM+CRF named entity recognition models. This technology supports the improvement of text analysis and knowledge mining capabilities in the field of tobacco.